PageIndex: Vectorless, Reasoning-based RAG

Summary

VectifyAI open-source vectorless RAG: build a JSON table-of-contents tree (node_id, title, page range, summary, sub_nodes) and let the LLM reason which section to read next, following in-document references instead of embedding similarity. Aimed at long structured docs (SEC filings, legal, manuals). Mafin 2.5, a PageIndex-powered financial RAG, reports 98.7% on FinanceBench (full set) vs typical vector RAG ~30–50% in the authors’ writeups. Ecosystem: OpenKB, ChatIndex, ConDB, pageindex-mcp; hosted chat/API at pageindex.ai. MIT. Discord post is a third-party Simplifying AI recap, not the authors.

Keywords

pageindex · vectorless-rag · vectify · financebench · mafin · tree-index · x

Topics

RAG, document retrieval, agentic search

Research notes

• Primary: GitHub README + API (MIT, Python, 35182 stars / 3092 forks at check). Discord/X https://x.com/simplifyinAI/status/2080468063548543454 via fxtwitter. Intro blog https://pageindex.ai/blog/pageindex-intro (cite zhang2025pageindex). FinanceBench numbers from Mafin 2.5 https://pageindex.ai/blog/Mafin2.5 / github.com/VectifyAI/Mafin2.5-FinanceBench. Retrieval framework, not a new hosted corpus, so no datasets_local row.